← SkySwap

How SkySwap works

SkySwap estimates disruption risk — the chance a flight is cancelled or arrives 15 or more minutes late — for U.S. domestic flights. It combines twelve months of government performance records with live conditions on the day of travel, and it publishes its validation results on this page so you can judge how well the model actually performs.

Data sources

Source	Used for	Coverage
U.S. Bureau of Transportation Statistics (BTS) On-Time Performance	historical cancellations, delays, taxi times	June 2025 – May 2026 · 151,682 flight-route records · 7,090 routes · 356 airports
NOAA Aviation Weather Center (METAR)	live weather at departure, en route, and arrival	live government feed
FAA National Airspace System status	live ground stops and ground delay programs	live government feed
NOAA/NWS point forecast + aviation TAF	forecast outlook at departure and arrival for travel dates up to 7 days ahead	government forecast; display context only — never an input to the score

The risk model

Every flight gets a base score from its historical record:

score = cancelPct × 3.0 + delayPct × 0.45 + avgDelay × 0.30

For travel dates up to 7 days ahead, SkySwap also shows a forecast outlook at the departure and arrival airports (NOAA/NWS point forecast, plus the aviation TAF within its ~30-hour window). This is display context only: forecasts are not validated the way the historical model is, so they never move the score. Beyond 7 days no airport-level government forecast exists, and the app says so rather than inventing precision.

The score is clamped to the range 4–100. On the day of travel only, a live adjustment is added on top of the base score:

live adjustment = (taxi-out − 15) × 0.8 + weatherImpact × 0.25

Per-flight history is blended toward the route average with weight n/(n+10) (empirical-Bayes shrinkage), so thin samples cannot produce inflated scores — a flight with only a handful of observed operations leans mostly on its route’s overall record.

Scores map to three bands: Low < 35, Moderate 35–59, High ≥ 60.

To be plain about it: this is SkySwap’s own model. It is not an industry standard, and no airline, regulator, or standards body has endorsed it.

Validation

The scoring method is validated out-of-sample against the most recent published BTS month the model has never seen. In the current experiment, the shipped model (trained on the 12 months through May 2026) was evaluated on 606,676 U.S. domestic flights from June 2026 (99.9% coverage). The overall June disruption base rate was 27.2%.

Two notes so these figures can be read correctly. First, this experiment is reproducible: the scoring replay ships in the repository (validate_holdout.cjs) and runs the app’s exact formula against the public BTS file for June 2026 — a month entirely outside the shipped model’s training window. The shipped dataset itself is likewise re-derivable: verify_data.cjs re-tallies the raw BTS files from scratch, checks every flight, route, and airport entry the app ships, and prints the raw record counts behind any single flight’s numbers. When the model is next rebuilt and absorbs June, the experiment is re-run against the next published month and these tables are refreshed. Second, absolute disruption rates vary with the season — the June 2026 base rate (27.2%) is well above the 20.9% measured in the previous run of this experiment on April 2026 data — so the percentages below describe June specifically. What the experiment validates is the ordering and separation of the score bands, which held in both months.

Band	Disrupted	Cancelled
Low <35	25.6%	1.3%
Moderate 35–59	43.0%	5.0%
High ≥60	49.9%	9.0%

Score range	Flights	Disrupted	Cancelled
4–14	182,124	17.3%	0.6%
15–24	271,163	28.2%	1.3%
25–34	98,811	33.9%	2.6%
35–44	35,679	41.0%	4.4%
45–59	15,387	47.5%	6.5%
60–74	2,844	49.8%	8.5%
75–100	668	50.0%	11.1%

Disruption rises monotonically with score across every range in the table. Flights scored High were disrupted at 1.8× the June base rate and cancelled at roughly 7× the rate of Low-scored flights.

What the data does not cover

The BTS On-Time Performance dataset is a legally mandated near-census, not a sample — but its scope is defined by regulation, and that scope has edges worth stating plainly.

Fourteen carriers only. Under 14 CFR 234.2, a “reporting carrier” is one accounting for at least 0.5% of domestic scheduled-passenger revenue. For 2026 that is 14 carriers. Operators such as Sun Country, Breeze, Avelo, Silver, Cape Air and Contour are absent entirely — SkySwap returns “not enough data” for them rather than guessing.
U.S. domestic only. No international segments, no foreign carriers, no cargo or charter operations.
Delay is measured against airline-set schedules. Scheduled block times are chosen by the carrier with no government oversight, and published research finds they have been padded by roughly 6–10 minutes per flight since the 1980s. A score therefore reflects scheduling policy as well as operational risk.
Cancellation has a narrow legal definition — a flight listed in the reservation system within seven days of departure and not operated. Schedule cuts made earlier do not appear as cancellations.
Publication lag of roughly two months, so the most recent weeks are never in the historical base.

Known limitations of the model

Only the historical base score is backtested. The live travel-day adjustment (taxi-out and weather) is not yet validated.
The model has no seasonality or time-of-day factors yet.
When a specific flight has no history, route-average-only estimates are used, and these discriminate weakly compared to flight-level history.
The gate-wait model assumes a fixed 40-minute minimum turn regardless of aircraft type.
Alternative flights shown are historical operators of a route, not verified schedules for the chosen date. Where a departure time or duration appears, it is that flight number’s typical scheduled value over the data window (the median of its BTS records), labelled “usually” — check the airline for the schedule on your date. Durations marked ≈ are estimates from route distance.
SkySwap must not be used for passenger-rights or compensation claims.

Independent assessment: the U.S. DOT Inspector General audited this dataset in 2024 (report AV2025003) and confirmed BTS verifies its accuracy through 117 automated checks, while noting that completeness and consistency checks could be strengthened. DOT relies on the same filings to penalise airlines for chronically delayed flights.